Reading Patterns, Not Just Inflation
A GWAS can produce test statistics larger than expected under a null model for several different reasons. Some variants may be associated with the phenotype because many genetic effects are real and collectively polygenic. Alternatively, confounding or a poorly calibrated test may inflate statistics even when the tested variants have no corresponding effects. These possibilities matter because the same overall inflation can imply very different conclusions about the evidence and the analysis.
A QQ plot, the genomic inflation factor \(\lambda\), and LD score regression each summarize a different feature of the association results. A QQ plot shows where departures occur across the distribution. Lambda compresses the distribution to one median ratio. LD score regression asks whether average test-statistic size changes with the amount of linkage disequilibrium tagging nearby variants. Their joint pattern can suggest explanations, but it does not by itself prove a particular cause.
For a quantitative phenotype in a simplified setting, the expected relationship under polygenic association is often written \(E(\chi_j^2) \approx 1 + (N h_g^2/M)\ell_j + a\). Here \(N\) is sample size, \(h_g^2\) is SNP heritability, \(M\) is the number of variants contributing to the model, and \(a\) represents an additive inflation component not explained by the modeled polygenic signal. Thus, the idealized intercept is \(1+a\), while polygenic signal contributes to the slope. This is a working diagnostic model, not a guarantee that every source of bias can be separated cleanly.
Interpret the summaries under suitable conditions: the association statistics should be calibrated under the null apart from the effects being investigated; the LD scores should be appropriate for the analyzed variants and ancestry; and the regression should use a suitable set of approximately independent markers with enough variation in LD score. Poor reference-panel matching, correlated errors, unusual variant filtering, or model misspecification can distort the intercept and slope. The course’s earlier material on LDSC assumptions and reference-panel mismatch explains why these checks matter.
How the Three Summaries Work Together
A QQ plot compares ordered observed p-values with the p-values expected under the null. If the null is well calibrated, most points should follow the reference line. A broad upward shift beginning among relatively ordinary p-values suggests widespread inflation. A curve that follows the line over most of its range and rises mainly at the extreme tail is more compatible with a substantial collection of null-like tests plus a subset of strong associations. That tail pattern is consistent with polygenicity, but it is not proof of it: a few technical or analytical problems can also create extreme results.
Lambda measures the median test statistic relative to the null median. It is easy to compare across analyses, but it does not show whether the shift is concentrated in the tail or spread across the distribution. It also depends on the number of tested variants and the amount of genuine signal. Therefore, a lambda above one is not, by itself, evidence of confounding or failed calibration.
LD score regression adds a different comparison: are larger statistics associated with larger LD scores? A positive slope can arise when causal variants are tagged by SNPs in higher-LD regions, as expected under the simplified polygenic model. An elevated intercept with little slope can be consistent with inflation that is not explained by LD-dependent tagging. Neither pattern uniquely establishes a cause. For example, some forms of confounding can correlate with LD score, and calibration problems may affect both the intercept and slope.
A Short Derivation of the LDSC Pattern
The diagnostic interpretation comes from the mean model. Suppose the expected association statistic for variant \(j\) can be approximated by a baseline of one, an additive background component \(a\), and an LD-dependent polygenic component:
Rearranging into a straight-line regression on \(\ell_j\) gives an intercept \(c=1+a\) and slope \(b=N h_g^2/M\). If there is no background inflation and no polygenic signal, the idealized values are \(c=1\) and \(b=0\). If there is polygenic signal under the model, the expected statistic increases with LD score and the slope is positive. If an additive source of inflation shifts statistics without that LD relationship, the intercept rises while the slope need not rise.
Now consider a simple global calibration error that multiplies the expected statistics by a factor \(q\). The expected regression becomes \(q(1+a)+q(Nh_g^2/M)\ell_j\). Both intercept and slope are multiplied by \(q\). This shows why a raised LDSC intercept is not a unique fingerprint of confounding: a global scaling error can also raise it. Evidence from null simulations, known negative controls, or a corrected analysis is needed to investigate calibration directly.
Three Simulated Diagnostic Case Studies
Worked Example: Inflation Mostly Unrelated to LD Score
Hypothetical simulation. A study simulates a phenotype with no causal genetic effects but introduces a source of confounding. Among one-degree-of-freedom tests, the observed median \(\chi^2\) statistic is \(0.5277\). LD score regression reports an intercept of \(1.120\) with standard error \(0.025\), and a slope of \(0.006\) with standard error \(0.004\). The QQ plot rises above its reference line across much of the distribution.
The inflation factor is \(\lambda=0.5277/0.4549=1.160\). The intercept is \(1.120\), about \(0.120/0.025=4.8\) standard errors above one. By contrast, the slope is \(0.006/0.004=1.5\) standard errors above zero, so these summaries do not provide strong evidence of a positive LD-dependent trend. The fitted mean at LD score \(2\) is \(1.120+0.006(2)=1.132\); at LD score \(20\), it is \(1.120+0.006(20)=1.240\).
The broad QQ departure, elevated lambda, and comparatively elevated intercept with a weak slope are consistent with the simulated confounding pattern. They do not identify which confounder produced it. A positive slope is also not ruled out: its estimate has uncertainty, and confounding can sometimes track LD score. The simulation’s known construction supports the interpretation here; in an observational analysis the diagnostics alone would not establish that explanation.
Worked Example: Tail Enrichment With an LD-Dependent Slope
Hypothetical simulation. A second phenotype has many small causal effects and no deliberately introduced confounding. The QQ plot follows the null reference through most of the distribution, then bends upward in its extreme tail. The observed median statistic is \(0.5459\). LD score regression gives an intercept of \(1.010\) (standard error \(0.020\)) and a slope of \(0.025\) (standard error \(0.004\)).
Here, \(\lambda=0.5459/0.4549=1.200\). The intercept is only \((1.010-1)/0.020=0.5\) standard errors above one, while the slope is \(0.025/0.004=6.25\) standard errors above zero. The fitted mean statistic at LD score \(2\) is \(1.010+0.025(2)=1.060\); at LD score \(20\), it is \(1.010+0.025(20)=1.510\). The increasing fitted mean illustrates the LD-dependent pattern captured by the slope.
Together, the tail-focused QQ departure, the slope clearly above zero, and the intercept close to one are consistent with polygenic signal under the simplified model. Lambda is elevated because real signals affect the statistic distribution, but its value alone cannot distinguish this case from the first example. In practice, an intercept close to one is not a certificate that confounding is absent; its estimate and the assumptions behind the regression still matter.
Worked Example: A Global Calibration Error That Resembles Confounding
Hypothetical simulation. A third analysis has a known calibration defect: test statistics are multiplied by \(q=1.10\). Before this defect, the simulated mean model is \(E(\chi_j^2)=1+0.010\ell_j\). After scaling, the expected model is \(1.10+0.011\ell_j\), since \(1.10(1)=1.10\) and \(1.10(0.010)=0.011\). Suppose the fitted regression reports intercept \(1.10\) and slope \(0.011\), and simulated null variants have median statistic \(0.5004\).
The lambda calculation is \(0.5004/0.4549=1.100\). At LD score \(10\), the fitted mean is \(1.10+0.011(10)=1.210\). The intercept is clearly above its ideal null value, but the slope has also been scaled. If only the regression were available, the high intercept might be mistaken for confounding, and a nonzero slope might be attributed to polygenic signal. The known simulation and the null-variant check instead reveal the global calibration defect.
This example does not justify automatically dividing every statistic by the fitted intercept. A correction should follow investigation of the model or implementation that caused miscalibration. Rescaling without that diagnosis can alter genuine signal and may not fix errors that vary across variants.
Diagnostics, Alternatives, and Failure Modes
When summaries disagree, treat disagreement as a prompt to inspect assumptions rather than a reason to select the most convenient statistic. Check the QQ plot at multiple portions of the distribution, inspect the uncertainty around the LDSC intercept and slope, and verify that the LD reference and analyzed variant set are suitable. If available, compare the observed statistics with known null simulations or negative controls run through the same analysis pipeline. Such checks are particularly useful for separating a genuine association pattern from broad test miscalibration.
| Pattern in the simulated summaries | Interpretation supported by the pattern | Important qualification |
|---|---|---|
| Broad QQ departure; intercept above one; slope small or uncertain | Consistent with inflation not explained by an LD-dependent polygenic trend | Does not identify the confounder; some confounding can affect the slope |
| QQ line mostly follows the reference before an upward tail; intercept near one; slope positive | Consistent with polygenic signal under the simplified LDSC model | Tail shape and slope are not proof of causality or correct calibration |
| Broad inflation, elevated intercept, and known null controls also inflated | Supports a calibration problem in the tested procedure or data handling | LDSC alone cannot diagnose the technical source or guarantee a valid correction |
Alternatives answer different questions. Lambda is a compact scalar and is useful for a quick comparison, but it loses distributional shape and LD dependence. A QQ plot makes the distributional shape visible, but does not summarize how statistics vary with LD score. LD score regression quantifies that relationship, but depends on its regression assumptions and reference data. Null simulations and negative controls can directly probe calibration in a defined setting, but their value depends on whether the simulated or control setting represents the analysis being assessed. These approaches complement one another rather than replacing one another.
A common mistake is to reason, “Lambda is greater than one, so the GWAS must be confounded.” That conclusion does not follow: polygenic signal can raise lambda, and calibration errors can do so as well. Another mistake is to treat an LDSC intercept above one as a direct measurement of confounding. The intercept is a fitted diagnostic quantity under a model, not a unique causal decomposition. The correction is to describe what the summaries show, state which explanations are consistent with them, and use independent checks before attributing a cause.
These diagnostics have limits even when calculated correctly. A single intercept and slope compress variant-level structure; heterogeneous effects, imperfect LD scores, sample overlap, and bias correlated with LD can complicate their interpretation. A high slope is not itself an estimate of the effect at an individual locus, and a small intercept estimate does not demonstrate that every association test is calibrated. Diagnostic evidence should be reported as evidence, not as certainty.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A simulated analysis has \(\lambda=1.18\), an LDSC intercept of \(1.02\), and a slope clearly above zero. Which explanation is consistent with this pattern, and what does lambda alone fail to establish?
- A QQ plot departs from the reference line across most of its range, while the fitted intercept is above one and the slope is imprecise. What is a cautious interpretation?
- Under a global multiplier \(q=1.05\), an unscaled mean model is \(1+0.020\ell_j\). What intercept and slope does the scaled model predict?
- Why is an LDSC intercept above one insufficient to prove that population confounding caused inflation?
Solutions. (1) The pattern is consistent with LD-dependent polygenic signal: the positive slope and near-one intercept support that interpretation under the model. Lambda alone establishes neither polygenicity nor confounding. (2) The pattern is consistent with broad inflation not clearly explained by the fitted LD-dependent trend, but it does not identify the cause; calibration checks and assumption review are needed. (3) Multiplication gives \(1.05(1+0.020\ell_j)=1.05+0.021\ell_j\), so the intercept is \(1.05\) and the slope is \(0.021\). (4) Other processes, including global calibration error, can raise the intercept, and model or reference-data problems can distort its estimate. The intercept is diagnostic evidence, not a unique causal label.